Note: This is the preliminary code of an ecommerce project done for an online vender. The Dataset, final code and report is confidential.

Robin Khatri

setwd("~/Desktop/E-Commerce")

data <- read.csv("ecommerce_dataset.csv", sep = ";")

library(dplyr)

Attaching package: ‘dplyr’

The following objects are masked from ‘package:stats’:

    filter, lag

The following objects are masked from ‘package:base’:

    intersect, setdiff, setequal, union

We'll first see a glimpse/summary of the dataset.

glimpse(data) # To see the variables and their types

Observations: 400,637
Variables: 13
$ Timestamp            <int> 1460462764, 1490572501, 1490572501, 1490572501...
$ User.ID              <fct> NULL, 0c6205543f3f6a7e8082, 0c6205543f3f6a7e80...
$ Cross.device.user.ID <fct> NULL, NULL, NULL, NULL, NULL, NULL, NULL, c05f...
$ Existing.client      <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0...
$ Event.type           <fct> Basket, Listing, Listing, Listing, Listing, Li...
$ Product.ID           <fct> Unknown, Unknown, Unknown, Unknown, Unknown, U...
$ Product.Price        <fct> 11, NULL, NULL, NULL, NULL, NULL, NULL, NULL, ...
$ Product.quantity     <fct> 1, NULL, NULL, NULL, NULL, NULL, NULL, 1, NULL...
$ Product.category     <fct> Medium Items, Packages, Packages, Packages, Pa...
$ Environment          <fct> web, web, web, web, web, web, web, web, web, w...
$ Device.type          <fct> Desktop, Desktop, Desktop, Desktop, iPhone, iP...
$ Browser.family       <fct> safari, edge, edge, edge, mobile safari, mobil...
$ User.location        <fct> Inland, Outside, Outside, Outside, Inland, Inl...

Questions¶

In this project, I have tried to answer the following questions from this dataset -

What is the evolution of the revenue during the year? Can you pinpoint and explain particular events during the year?
Over the year, what devices have shoppers used to browse this merchant's site? Explain the main changes.
In your opinion, do Applications perform better than Mobile Web browsing?
Are there any changes in the website during the year and is yes, what is its impact?
What is the difference in customer behavior and what is the rate of customer aquisition?
Possibility of customer segmentation

Points to note¶

Variables User.location and Browser.family were not considered in this analysis
Sales to Visits ratio refered in this presentation corresponds to "Number of Sales/Number of visits to pages containing products of same kind"

Data Preparation¶

Date was converted into Date object from UNIX Timestamps
Month and Year vectors were generated from Date
Duplicate instances (rows) were removed
Data was ordered according to date and time as it wasn't ordered before.

About dataset¶

After preparing the data for analysis, there were 384841 instances of single events from 28 March 2016 to 17 April 2017

We'll look for duplicates in data and remove them.

nrow(distinct(data)) # Get distinct instances/rows

nrow(data) - nrow(distinct(data))

15796 instances are duplicates. We'll remove them from out dataset

data <- distinct(data)

Now we'll get date and times from timestamps.

library(lubridate)

Attaching package: ‘lubridate’

The following object is masked from ‘package:base’:

    date

data$Date <- as.POSIXct(data$Timestamp, origin="1970-01-01")

head(data$Date)

[1] "2016-04-12 14:06:04 CEST" "2017-03-27 01:55:01 CEST"
[3] "2017-03-27 01:38:06 CEST" "2017-03-27 01:22:51 CEST"
[5] "2017-03-27 01:00:49 CEST" "2017-03-27 01:00:29 CEST"

tail(data$Date)

[1] "2016-12-07 13:35:54 CET" "2016-12-07 13:39:41 CET"
[3] "2016-12-07 12:30:11 CET" "2017-02-08 10:32:22 CET"
[5] "2017-02-08 10:27:08 CET" "2017-01-25 10:21:47 CET"

The dataset is not ordered by date. We'll do so.

data <- data[order(data$Date),]

head(data$Date); tail(data$Date)

[1] "2016-03-28 02:17:38 CEST" "2016-03-28 02:30:53 CEST"
[3] "2016-03-28 02:34:44 CEST" "2016-03-28 03:10:31 CEST"
[5] "2016-03-28 03:21:53 CEST" "2016-03-28 03:22:50 CEST"

[1] "2017-04-17 00:45:17 CEST" "2017-04-17 00:45:26 CEST"
[3] "2017-04-17 00:45:40 CEST" "2017-04-17 00:53:18 CEST"
[5] "2017-04-17 00:55:07 CEST" "2017-04-17 01:18:13 CEST"

Now our data is ordered by time. CEST represents Central European Summer Time.

min(data$Date); max(data$Date)

[1] "2016-03-28 02:17:38 CEST"

[1] "2017-04-17 01:18:13 CEST"

data$Month <- month(data$Date, label=TRUE); head(data$Month)
data$Year <- year(data$Date); head(data$Date)
data$weekday <- wday(data$Date, label=TRUE); head(data$weekday)
data$time <- format(data$Date, "%H:%M:%S"); head(data$time)
data$hour <- hour(data$Date); head(data$hour)
data$YearMonth <- paste(data$Year, data$Month); head(data$YearMonth) # To be clear which month of which year

[1] "2016-03-28 02:17:38 CEST" "2016-03-28 02:30:53 CEST"
[3] "2016-03-28 02:34:44 CEST" "2016-03-28 03:10:31 CEST"
[5] "2016-03-28 03:21:53 CEST" "2016-03-28 03:22:50 CEST"

So, we have data available from 28 March 2016 to 17 April 2017

Now, we'll convert Columns Product.Price and Product.quantity into numerical ones. Currently they are factors.

head(data$Product.Price)

data$Price <- as.numeric(levels(data$Product.Price)[data$Product.Price]); head(data$Price)

Warning message in eval(expr, envir, enclos):
“NAs introduced by coercion”

data$Quantity <- as.numeric(levels(data$Product.quantity)[data$Product.quantity]); head(data$Quantity)

Warning message in eval(expr, envir, enclos):
“NAs introduced by coercion”

NA's will be introduced by coercion inplace of "NULL" level of the original factor.

data$Revenue <- data$Price*data$Quantity; head(data$Revenue)

Now we'll get data specific to sales.

levels(data$Event.type) # To get levels of Events. There should be four as per guidelines

sdata <- data[data$Event.type=="Sales",]# Number of rows in sales data
paste("Total number of transactions is:", nrow(sdata))

sales.prop <- nrow(sdata)*100/nrow(data); round(sales.prop,2) # Proportion of sales data of total data

Only 5.86 % of the data is of sales

Getting total revenue

revenue16.17 <- sum(sdata$Revenue); paste("Total revenue is:", revenue16.17)

This is the total revenue in the period 28 March 2016 to 17 April 2017.

Now, we'll try to see how revenue changed over time.

library(ggplot2)

Revenue: Evolution through the year¶

options(repr.plot.width=8, repr.plot.height=3)
sdata %>%
  group_by(Date) %>%
  summarise(r = sum(Revenue)) %>%
  ggplot(aes(x = Date, y = r)) + geom_line() + geom_smooth(method = 'auto', se = FALSE) + labs(x = 'Date', y = 'Revenue (Euro)', title = 'Revenue by Date')

`geom_smooth()` using method = 'gam' and formula 'y ~ s(x, bs = "cs")'

We'll see the location of the shoppers.¶

summary(sdata$User.location) # Number of transactions by User's location
x <- sum(sdata$Revenue[sdata$User.location=="Inland"]) # Revenue
y <- sum(sdata$Revenue[sdata$User.location=="Outside"])
paste("Revenue generated from Inland Users is: ", x)
paste("Revenue generated from Outside Users is: ", y)

Sales to Visits (To product specific pages) is similar for Inland and Outside users.

# Create a dummy variable having 1's indicating Sales
sdata$dummy[sdata$Event.type=="Sales"] <- 1

options(repr.plot.width=8, repr.plot.height=3)
sdata %>%
  group_by(Date) %>%
  summarise(t = sum(dummy)) %>%
  ggplot(aes(x = Date, y = t)) + geom_line() + geom_smooth(method = 'auto', se = FALSE) + labs(x = 'Date', y = 'Number of transactions', title = 'Number of transactions by Date')

`geom_smooth()` using method = 'gam' and formula 'y ~ s(x, bs = "cs")'

We'll get top 5 days for revenue and also transactions.

date.rev <- sdata %>%
    group_by(Date) %>%
    summarise(Rev. = sum(Revenue)) %>%
    arrange(desc(Rev.))
date.tra <- sdata %>%
    group_by(Date) %>%
    summarise(Tra = sum(dummy)) %>%
    arrange(desc(Tra))

topdate.rev <- top_n(date.rev,5, date.rev$Rev.); topdate.rev
topdate.tra <- top_n(date.tra,5, date.tra$Tra); topdate.tra

Table above shows instances of Sales with most revenue in a single day.

All of these values belong to phase Nov - Early March.
For December, November and January, high revenue generated is clearly due to high amount of transactions taking place during this period.
For 31 March and 14 April 2017, the reason for high amount of revenue is not higher number of transactions. We'll see all Revenue data from these top 5 Revenue dates.

p = list()
for (i in 1:nrow(topdate.rev)){
p[[i]] = sdata[sdata$Date==topdate.rev$Date[i],]$Revenue
}
names(p) = paste("Revenue generated from transaction(s) on ", topdate.rev$Date)
p

As it turns out, the spike for those individual days in April and March was due to a single expensive purchase from user IDs d926128c491174a3f549 and 6a7410d13d7dc7549bdb
We'll explore User ID based data later on in detail.
Now we know the reason of high revenue two days in April 2017 and March 2017.

q = list()
for (i in 1:2){
q[[i]] = sdata[sdata$Date==topdate.rev$Date[i],]$User.ID
}
q

So, now we know the reason of high revenue two days in April 2017 and March 2016.

To get a clear picture, we'll get revenue by month.¶

Considerations:¶

Observations of only 3 and 17 days available respectively for March 2016 and April 2017.
To compare monthly revenue, we'll remove March 2016 and April 2017 from our dataset.

sdata.new <- sdata[!sdata$YearMonth %in% c("2016 Mar", "2017 Apr"),]
month.rev <- sdata.new %>%
    group_by(YearMonth) %>%
    summarise(Rev. = sum(Revenue)) %>%
    arrange(desc(Rev.))
month.rev$Revenue.pct = round(100*month.rev$Rev./sum(month.rev$Rev.),2)

colnames(month.rev)[3] = "Percentage"

month.rev

month.rev$levels = factor(month.rev$YearMonth, levels= month.rev$YearMonth[c(7,9,11,12,10,5,6,2,1,4,3,8)])

ggplot(month.rev, aes(x = levels, y = Rev.)) + geom_col() + labs(x = 'Month', y = 'Revenue', title = 'Revenue') + geom_text(aes(label=Percentage),position=position_stack(vjust = 0.5))

Revenue by month¶

It is clear that most of the revenue (Around 50%) was generated during Nov 2016 - Feb 2017.
Since the dataset is from retail, this behavior maybe due to seasonality present in data. It peaks during Nov-Dec due to holidays and increase in number of sales.
Thereafter it began to decline. The revenue generated during other months is very less.
In order to see this more clearly, we get number of transactions and average order value by month.

sdata.new <- sdata[!sdata$YearMonth %in% c("2016 Mar", "2017 Apr"),]
month.rev <- sdata.new %>%
    group_by(YearMonth) %>%
    summarise(Rev. = sum(Revenue), Tran = sum(dummy)) %>%
    mutate(AOV = round(Rev./Tran,2)) %>%
    arrange(desc(Rev.))
month.rev$Percentage <- round(month.rev$Tran*100/sum(month.rev$Tran),2)

month.rev
summary = list()
for(i in 1:3){
    summary[[i]] = c(round(mean(as.data.frame(month.rev)[,i+1]),2), round(sqrt(var(as.data.frame(month.rev)[,i+1])),2))}
names(summary) <- c("Avg & Stdev Revevenue", "Avg & Stdev Transactions", "Avg & Stdev AOV"); summary
month.rev$levels = factor(month.rev$YearMonth, levels= month.rev$YearMonth[c(7,9,11,12,10,5,6,2,1,4,3,8)])

ggplot(month.rev, aes(x = levels, y = AOV)) + geom_col() + labs(x = 'Month', y = 'Average Order Value', title = 'Average Order Value')

ggplot(month.rev, aes(x = levels, y = Tran)) + geom_col() + labs(x = 'Month', y = 'Number of Daily Transactions', title = 'Number of Transactions') + geom_text(aes(label=Percentage),position=position_stack(vjust = 0.5))

summary

The higher revenue during months of Nov-Dec is due to increase in number of sales.
The average order value is more or less same throughout the year.
Standard deviation for transactions over the month is nearly half of the average (853.13) while average order value is nearly the same.

We'll check statistical significance of difference in number of transactions.¶

library(agricolae)
monthlysummary = sdata %>%
  group_by(date(Date), Month) %>%
  summarise(Tran = n_distinct(Timestamp)) %>%
  ungroup()

kruskal.test(Tran ~ Month, data = monthlysummary)

	Kruskal-Wallis rank sum test

data:  Tran by Month
Kruskal-Wallis chi-squared = 162.12, df = 11, p-value < 2.2e-16

kruskal(monthlysummary$Tran, monthlysummary$Month, console = TRUE)

Study: monthlysummary$Tran ~ monthlysummary$Month
Kruskal-Wallis test's
Ties or no Ties

Critical Value: 162.1231
Degrees of freedom: 11
Pvalue Chisq  : 0 

monthlysummary$Month,  means of the ranks

    monthlysummary.Tran  r
Apr           164.36957 46
Aug           109.22581 31
Dec           298.95161 31
Feb           288.01786 28
Jan           241.10000 30
Jul            90.85484 31
Jun           104.35000 30
Mar           150.01429 35
May           142.09677 31
Nov           285.11667 30
Oct           219.98387 31
Sep           247.75000 30

Post Hoc Analysis

t-Student: 1.966362
Alpha    : 0.05
Groups according to probability of treatment differences and alpha level.

Treatments with the same letter are not significantly different.

    monthlysummary$Tran groups
Dec           298.95161      a
Feb           288.01786     ab
Nov           285.11667     ab
Sep           247.75000     bc
Jan           241.10000      c
Oct           219.98387      c
Apr           164.36957      d
Mar           150.01429     de
May           142.09677    def
Aug           109.22581    efg
Jun           104.35000     fg
Jul            90.85484      g

As suspected the difference in number of transactions over months is statistically significant.

To find paiwise significance between tests, we ran post-hoc test and found following results -

Groups with the same letter are not significantly different.

Month	Letter
Jan	c
Feb	ab
Mar	de
Apr	d
May	def
Jun	fg
Jul	g
Aug	efg
Sep	bc
Oct	c
Nov	ab
Dec	a

Now we'll look into Visits to sales ratio. For this purpose we'll consider only visits to product specific webpages.

# Get product visits
pdata <- data[data$Event.type=="Product",]
paste("No. of visits to product specific web-pages of the website is:", nrow(pdata))

# Get product visits by month
# Remove April 2017 and March 2016

# Create a dummy variable having 1's 
pdata$dummy <- rep(0,nrow(pdata))
pdata$dummy[pdata$Event.type=="Product"] <- 1

pdata.new <- pdata[!pdata$YearMonth %in% c("2016 Mar", "2017 Apr"),]
month.prod <- pdata.new %>%
    group_by(YearMonth) %>%
    summarise(Visits = sum(dummy)) %>%
    ungroup()

month.prod$sales <- rep(0,12)
month.prod$svisits <- rep(0,12)

for(i in 1:12){
    month.prod$sales[i] <- month.rev$Tran[which(month.rev$YearMonth == month.prod$YearMonth[i])]
}

for(i in 1:12){
month.prod$svisits[i] <- round(month.prod$sales[i]/month.prod$Visits[i], 2)
}

month.prod[order(-month.prod$svisits),]

It is clear that there is an inconsistency when it comes to user purchases after viewing the product.
Four out of top 5 months by revenue are also the top 5 months by sales to visits ratio.
September onwards the sales to visit ratio has been better.
Consider for instance, in Feb 2017 and May 2016, the difference in sales is more than 1000 but number of visits to product specific pages is nearly the same.
An idea to increase the number of sales (and so the revenue) is to focus on converting these visits to sales.
This will result in increas in revenue during months in which revenue was less.

We'll now consider the devices which users used to access the website of the retailer and to make purchases.¶

Devices -¶

levels(factor(data$Device.type))

Environment -¶

levels(data$Environment)

Let us see which environment was used in which device.

t(table(data$Environment, data$Device.type))

                      
                       app_android    web
  Android - Smartphone           0  12213
  Android - Tablet               0   5138
  Desktop                        0 286511
  iPad                           0  12423
  iPhone                         0  22678
  Mobile - Other                 0   4597
  Unknown                    41281      0

It seems that the category "Unknown" of the devices is made up of app_android environment.

So, It is necessary to know to evaluate the difference in performance of mobile web browsing and app_android.

First we'll look at the usage of devices over the year. For this purpose, we'll divide the devices into two distinct categories:

1. Mobile (Android and ios smartphones and other phones, and Unknown)¶

2. Desktop¶

3. Tablets (Android and ios tablets)¶

Total 2016-17 revenue by device type¶

data$Device <- rep(0, nrow(data))
data$Dummy2 = data$Dummy1 = data$Device = data$Dummy
data$Device.type = as.character(data$Device.type)
data$Device[data$Device.type %in% c("Android - Smartphone", "Android - Tablet", "iPhone" , "Mobile - Other", "Unknown")] = "Mobile"
data$Device[data$Device.type == "Desktop"] = "Desktop"
data$Device[data$Device.type %in% c("iPad", "Android - Tablet")] = "Tablet"
data$Dummy[data$Event.type == "Sales"] = 1
data$Dummy1[data$Event.type == "Product"] = 1
data$Dummy2[data$Event.type == "Basket"] = 1
data$Device <- factor(data$Device)
levels(data$Device)

Revenuebydevices <- data[data$Dummy==1,] %>%
    group_by(Device) %>%
    summarise(Revenue = sum(Revenue, na.rm=TRUE), Tran = sum(Dummy)) %>%
    mutate(AOV = round(Revenue/Tran,2), Revenue.pct = round(Revenue / sum(Revenue) * 100, 2)) %>%
    arrange(desc(Revenue))

Visitsbydevices <- data[data$Dummy1==1,] %>%
    group_by(Device) %>%
    summarise(Visits = sum(Dummy1))

Basketsbydevices <- data[data$Dummy2==1,] %>%
    group_by(Device) %>%
    summarise(Baskets = sum(Dummy2), Basketvalue = sum(Revenue)) %>%
    mutate(ABV = round(Basketvalue/Baskets,2))
Revenuebydevices$svisits <- round(Revenuebydevices$Tran/Visitsbydevices$Visits,2); Revenuebydevices$ABV <- Basketsbydevices$ABV; Revenuebydevices$Product.visits <- Visitsbydevices$Visits

Revenuebydevices <- Revenuebydevices[complete.cases(Revenuebydevices),]; Revenuebydevices
ggplot(Revenuebydevices, aes(x="", y=Revenue/1000, fill=Device))+ ylab("Revenue (in 1000's)") + geom_bar(width = 1, stat = "identity") + coord_polar("y", start=0) + scale_fill_brewer(palette="Blues")+ theme_minimal()

Around 77 % of the revenue was generated from Desktop while close to 22 % was generated on Mobile.
Average basket value was almost consistent with the Average order value.
Looking at the sales to visits (To product specific pages) ratio, it is clear that for tablets the ratio was much lower which means that on tablets people proportion of visitors who make a sale is quite less (2%).

We'll see if there is any statistically significant different between devices when it comes to generating Revenue and their sales to visits ratio.

Since this is a count data and the categories: Products and Transactions may not be indenpendent. So, we used Fisher's exact test instead of Chi Square Test.

M <- as.matrix(cbind(Revenuebydevices$Tran, as.factor(Revenuebydevices$Product.visits)))
fisher.test(M, y = Revenuebydevices$Device)

	Fisher's Exact Test for Count Data

data:  M
p-value = 0.05802
alternative hypothesis: two.sided

kruskal.test(Revenue ~ as.factor(Device), data = data[data$Event.type=="Sales",])
# Excluding tablets to compare Desktop and Mobile

kruskal(data[data$Event.type=="Sales",]$Revenue, as.factor(data[data$Event.type=="Sales",]$Device), console = TRUE)

	Kruskal-Wallis rank sum test

data:  Revenue by as.factor(Device)
Kruskal-Wallis chi-squared = 481.17, df = 2, p-value < 2.2e-16

Study: data[data$Event.type == "Sales", ]$Revenue ~ as.factor(data[data$Event.type == "Sales", ]$Device)
Kruskal-Wallis test's
Ties or no Ties

Critical Value: 481.1672
Degrees of freedom: 2
Pvalue Chisq  : 0 

as.factor(data[data$Event.type == "Sales", ]$Device),  means of the ranks

        data.data.Event.type.....Sales.....Revenue     r
Desktop                                   11855.49 16314
Mobile                                     9712.63  5900
Tablet                                    10114.46   324

Post Hoc Analysis

t-Student: 1.960069
Alpha    : 0.05
Groups according to probability of treatment differences and alpha level.

Treatments with the same letter are not significantly different.

        data[data$Event.type == "Sales", ]$Revenue groups
Desktop                                   11855.49      a
Tablet                                    10114.46      b
Mobile                                     9712.63      b

For Mobile and Desktop, the sales to visits (Visits to product specific pages) is not significant.¶

When it comes to user aquisition, both desktop and Mobile perform equally well.

Difference in generated revenue between Mobiles and Desktops is statistically significant. More revenue can be generated by increasing the number of visits to the retailer's website since the sales to product visits ratio is not statistically significant among devices.

Now we'll look at the distribution of Revenue generated, Number of transactions and Average order value throughout the year for individual device type.

Mobile¶

Mobile <- data[data$Device.type %in% c("Android - Smartphone", "Android - Tablet", "iPhone" , "Mobile - Other", "Unknown"),]
data.new <- Mobile[!Mobile$YearMonth %in% c("2016 Mar", "2017 Apr"),]
dataforplot <- data.new[data.new$Dummy==1,] %>%
    group_by(YearMonth) %>%
    summarise(Revenue = sum(Revenue, na.rm=TRUE), Tran = sum(Dummy, na.rm =TRUE)) %>%
    mutate(AOV = round(Revenue/Tran,2))
dataforplot <- dataforplot[complete.cases(dataforplot),]
dataforplot$Percentage <- round(dataforplot$Revenue*100/sum(dataforplot$Revenue),2)
dataforplot$Percentage1 <- round(dataforplot$Tran*100/sum(dataforplot$Tran),2)
dataforplot$Percentage2 <- round(dataforplot$AOV*100/sum(dataforplot$AOV),2)
dataforplot$levels = factor(dataforplot$YearMonth, levels= dataforplot$YearMonth[c(1, 6,5,4,2,9,8,7,3,11,10,12)])

temp = list(); temp[[1]] <- c(round(mean(dataforplot$Tran),2), round(sqrt(var(dataforplot$Tran)),2)); names(temp) <- "Avg and Stdev Tran"; temp
temp = list(); temp[[1]] <- c(round(mean(dataforplot$Revenue),2), round(sqrt(var(dataforplot$Revenue)),2)); names(temp) <- "Avg and Stdev Revenue"; temp
temp = list(); temp[[1]] <- c(round(mean(dataforplot$AOV),2), round(sqrt(var(dataforplot$AOV)),2)); names(temp) <- "Avg and Stdev AOV"; temp

The average order value per month through mobile (16.55 Euros)(Including applications) was less than the average order value per month overall (Around 20 Euros).
The variation in average order value for Mobile is more or less comparable to the one we saw for overall Average order value

ggplot(dataforplot, aes(x = levels, y = Revenue)) + geom_col() + labs(x = 'Month', y = 'Total Revenue', title = 'Revenue (Mobile)') + geom_text(aes(label=Percentage),position=position_stack(vjust = 0.5))

ggplot(dataforplot, aes(x = levels, y = Tran)) + geom_col() + labs(x = 'Month', y = 'Number of Transactions', title = 'Number of Transactions (Mobile)') + geom_text(aes(label=Percentage1),position=position_stack(vjust = 0.5))

ggplot(dataforplot, aes(x = levels, y = AOV)) + geom_col() + labs(x = 'Month', y = 'Average Order Value', title = 'Average Order Value (Mobile)')

Something happened in September 2016 which caused change in pattern of number of transactions generation thorugh Mobile.¶

We'll soon look into this when we look transactions through Mobile only.¶

Desktop¶

Desktop <- data[data$Device.type == "Desktop",]
data.new <- Desktop[!Desktop$YearMonth %in% c("2016 Mar", "2017 Apr"),]
dataforplot <- data.new[data.new$Dummy==1,] %>%
    group_by(YearMonth) %>%
    summarise(Revenue = sum(Revenue, na.rm=TRUE), Tran = sum(Dummy)) %>%
    mutate(AOV = round(Revenue/Tran,2))
dataforplot <- dataforplot[complete.cases(dataforplot),]
dataforplot$levels = factor(dataforplot$YearMonth, levels= dataforplot$YearMonth[c(1, 6,5,4,2,9,8,7,3,11,10,12)])

temp = list(); temp[[1]] <- c(round(mean(dataforplot$Tran),2), round(sqrt(var(dataforplot$Tran)),2)); names(temp) <- "Avg and Stdev Desktop Transactions"; temp
temp = list(); temp[[1]] <- c(round(mean(dataforplot$Revenue),2), round(sqrt(var(dataforplot$Revenue)),2)); names(temp) <- "Avg and Stdev Desktop Revenue"; temp
temp = list(); temp[[1]] <- c(round(mean(dataforplot$AOV),2), round(sqrt(var(dataforplot$AOV)),2)); names(temp) <- "Avg and Stdev Desktop AOV"; temp

The average order value for Desktop has been better than mobile.

ggplot(dataforplot, aes(x = levels, y = Revenue)) + geom_col() + labs(x = 'Month', y = 'Total Revenue', title = 'Revenue (Desktop)')

ggplot(dataforplot, aes(x = levels, y = Tran)) + geom_col() + labs(x = 'Month', y = 'Number of Daily Transactions', title = 'Number of Transactions (Desktop)')

ggplot(dataforplot, aes(x = levels, y = AOV)) + geom_col() + labs(x = 'Month', y = 'Average Order Value', title = 'Average Order Value (Desktop)')

Excluding "Unknown" from Mobile, i.e. considering only Mobile Web Browsing¶

Mobile.excl.unknown <- data[data$Device.type %in% c("Android - Smartphone", "Android - Tablet", "iPhone" , "Mobile - Other"),]
data.new <- Mobile.excl.unknown[!Mobile.excl.unknown$YearMonth %in% c("2016 Mar", "2017 Apr"),]
dataforplot <- data.new[data.new$Dummy==1,] %>%
    group_by(YearMonth) %>%
    summarise(Revenue = sum(Revenue, na.rm=TRUE), Tran = sum(Dummy, na.rm=TRUE)) %>%
    mutate(AOV = round(Revenue/Tran,2))
dataforplot <- dataforplot[complete.cases(dataforplot),]
dataforplot$levels = factor(dataforplot$YearMonth, levels= dataforplot$YearMonth[c(1, 6,5,4,2,9,8,7,3,11,10,12)])

temp = list(); temp[[1]] <- c(round(mean(dataforplot$Tran),2), round(sqrt(var(dataforplot$Tran)),2)); names(temp) <- "Avg and Stdev Tran"; temp
temp = list(); temp[[1]] <- c(round(mean(dataforplot$Revenue),2), round(sqrt(var(dataforplot$Revenue)),2)); names(temp) <- "Avg and Stdev Revenue"; temp

ggplot(dataforplot, aes(x = levels, y = Revenue)) + geom_col() + labs(x = 'Month', y = 'Total Revenue', title = 'Revenue (Mobile Web Browsing)')

ggplot(dataforplot, aes(x = levels, y = Tran)) + geom_col() + labs(x = 'Month', y = 'Number of Daily Transactions', title = 'Number of Transactions (Mobile Web Browsing)')
ggplot(dataforplot, aes(x = levels, y = AOV)) + geom_col() + labs(x = 'Month', y = 'Average Order Value', title = 'Average Order Value (Mobile Web Browsing)')

We see that total revenue generated from Mobile web browsing fluctuated quite a lot with a standard deviation of around 340 Euros.

We also saw that the number of transactions has been decreased when it was increased for desktop.

Therefore it is necessary to look at data generated through User Agent: App_Android (or device: Uknown)

App_android (Device category: "Unknown")¶

unknown <- data[data$Device.type == "Unknown",]
min(unknown$Date)
max(unknown$Date)

[1] "2016-09-01 09:39:07 CEST"

[1] "2017-04-17 00:55:07 CEST"

Application (App_android) data is available from 01 September. My guess is application was launched then.¶

data.new <- unknown[!unknown$YearMonth %in% c("2016 Mar", "2017 Apr"),]
dataforplot <- data.new[data.new$Dummy==1,] %>%
    group_by(YearMonth) %>%
    summarise(Revenue = sum(Revenue, na.rm=TRUE), Tran = sum(Dummy, na.rm = TRUE)) %>%
    mutate(AOV = round(Revenue/Tran,2))
dataforplot <- dataforplot[complete.cases(dataforplot),]

temp = list(); temp[[1]] <- c(round(mean(dataforplot$Tran),2), round(sqrt(var(dataforplot$Tran)),2)); names(temp) <- "Avg and Stdev Transactions through Android Application"; temp
temp = list(); temp[[1]] <- c(round(mean(dataforplot$Revenue),2), round(sqrt(var(dataforplot$Revenue)),2)); names(temp) <- "Avg and Stdev Revenue through Android Application"; temp

dataforplot$levels = factor(dataforplot$YearMonth, levels= dataforplot$YearMonth[c(4,3,2,1,6,5,7)])

ggplot(dataforplot, aes(x = levels, y = Revenue)) + geom_col() + labs(x = 'Month', y = 'Total Revenue', title = 'Revenue (Android Applications)')

ggplot(dataforplot, aes(x = levels, y = AOV)) + geom_col() + labs(x = 'Month', y = 'Average Order Value', title = 'Average Order Value (Android Applications)')

ggplot(dataforplot, aes(x = levels, y = Tran)) + geom_col() + labs(x = 'Month', y = 'Number of Daily Transactions', title = 'Number of Transactions (Android Applications)')

The average revenue generated from applications is almost 7 times the average revenue generated through mobile web browsing.
Same goes for transactions.
The average order value is not very much as we have already seen that average order value did not vary much.
Before the launch of the application, the number of transactions through mobile phones were less than 200 per month
After the launch, the number of transactions through phones are averaging 1000. So, it is fair to say that Applications perform better than mobile web browsing.
So, retailer should invest more in development of application to provide customer-centric support. :)

Visitors to listing page¶

listing <- data[data$Event.type == "Listing",]
paste("first listing date is: ", min(listing$Date), "and its range is: ", round((max(listing$Date)-min(listing$Date)),2))
listing.mobile <- listing[!listing$Device.type %in% c("Desktop", "Tablet"),] # Mobile
range = max(listing.mobile$Date)-min(listing.mobile$Date)
paste("first listing event type on mobile: ", min(listing.mobile$Date), "and its range is: ", round(range,2))
# For desktop
listing.desktop <- listing[listing$Device.type == "Desktop",] # Desktop
paste("first listing date on Desktop is: ", min(listing.desktop$Date), "and its range is: ", round(max(listing.desktop$Date)-min(listing.desktop$Date),2))

Before 22 March 2017, there was no page dedicated to products belonging to the same category.

My guess is that it was created during this period.
Possibly this maybe why the sales figures in April 2017 were better than April 2016. Although data from full month of April 2017 was not available.

We'll check if there has been any change in sales to visits ratio after inclusion (supposedly) of pages where products of similar kinds are listed.

listing1 <- data[data$Event.type %in% c("Listing", "Sales", "Product"), ]

listing1 <- listing1[listing1$Date >= min(listing.mobile$Date),]
listing1 <- listing1[listing1$Date <= max(listing.mobile$Date),]
sales <- nrow(listing1[listing1$Event.type=="Sales",])
visits <- nrow(listing1[listing1$Event.type %in% c("Listing", "Product"),])
svisits.listing <- sales/visits

round(svisits.listing, 2)

The sales to visits ratio for the period is 0.08 which is better than data during the same duration in 2016. But again, the sample is too small to comment on the impact of inclusion of pages containing grouped products

Event: Basket¶

listing <- data[data$Event.type == "Basket",]
paste("first listing date is: ", min(listing$Date), "and its range is: ", round((max(listing$Date)-min(listing$Date)),2))
# For mobile
listing.mobile <- listing[!listing$Device.type %in% c("Desktop", "Tablet"),] # Mobile
range = max(listing.mobile$Date)-min(listing.mobile$Date)
paste("first listing event type on mobile: ", min(listing.mobile$Date), "and its range is: ", round(range,2))
# For desktop
listing.desktop <- listing[listing$Device.type == "Desktop",] # Desktop
paste("first listing date is: ", min(listing.desktop$Date), "and its range is: ", round(max(listing.desktop$Date)-min(listing.desktop$Date),2))

In order to make our approach to market to the customers, we should have a way to:

Divide users into groups based on their purchasing habits and history on our website.
We also want a new customer who just came to our website and included items in our basket to be classified in a group.

Considerations -¶

We do not know Cross Device User ID of every customer.
For a new customer, Cross Device User ID will not be known.

Therefore, for better grouping of customers. We consider two approaches -

Divide the customers whose Cross Device User ID is known to us in simple groups which may provide better understanding them.
Cluster the customers by Cross Device User ID and treat the UserID of a new customer as Cross Device User ID until the Cross Device User ID is obtained. This is the best approach but it is the best possible approach since clustering through User IDs will lack robustness of results.

Customer segmenation (For users with known Cross Device User ID)¶

userdata <- sdata[sdata$Cross.device.user.ID!="NULL",]
paste("Number of User IDs available is: ", nrow(userdata))

We'll create a new category "Lcustomers" (Short for Loyal customers)

This category will cosist of those customers who have shopped with this retailer more than twice.

We shall also create two other categories -

ICustomers (Short for important customers - for users who have bought more than third quartile of the Average Order Value (AOV).

SCustomers (Short for sleeping customers - for users who purchase with us more than two months ago and have not purchased since.

Note that these categories are not mutually exclusive.

However, this information is vital if we want to target these customers depending on their behavior on the retailer's website.

First, we'll look at how many distinct customers are there, and how many and which items they have purchased. We'll also consider the total amount they spent, and days since their last purchase

userdata$package[userdata$Product.category=="Packages"] = 1
userdata$medium[userdata$Product.category=="Medium Items"] = 1
userdata$small[userdata$Product.category=="Small Items"] = 1
userdata$large[userdata$Product.category=="Large Items"] = 1
userdata$dummy[userdata$Event.type=="Sales"] = 1
userdata$date <- as.Date(userdata$Date)

users <- userdata %>%
    group_by(Cross.device.user.ID) %>%
    summarise(package = sum(package), medium = sum(medium), small=sum(small), large=sum(large), Tran = sum(dummy), Revenue = sum(Revenue), quantity = sum(Quantity, na.rm=TRUE), firstdate = min(date), lastdate = max(date)) %>%
    mutate(time.first = as.numeric(max(as.Date(sdata$Date)) - firstdate), time.last = as.numeric(max(as.Date(sdata$Date)) - lastdate))
users[is.na(users)] = 0

head(users) 
paste("Number of users with 1 transaction is: ", nrow(users[users$Tran==1,]))
paste("Number of users with more than 1 transactions is: ", nrow(users[users$Tran>1,]))
paste("Number of users with more than 2 transactions is: ", nrow(users[users$Tran>2,]))
paste("Maximum number of transaction for one user is: ", max(users$Tran))
paste("Avg number of transactions per user is is: ", round(mean(users$Tran),2))
paste("Maximum number of product quantities bought for one user is: ", max(users$quantity))
paste("Avg product quantities bought is: ", round(mean(users$quantity),2))
paste("For users, correlation between number of transactions and revenue is: ",round(cor(users$Revenue, users$Tran),2))
paste("Maximum difference between first and last date of purchase is: ", max(users$time.first-users$time.last))
vector <- c(sum(users$package), sum(users$medium), sum(users$small), sum(users$large))
names <- c("package", "medium", "small", "large")
paste("Product type bought most is: ", names[which(vector==max(vector))])

Category: LCustomers (Loyal Customers)¶

Lcustomers <- users[users$Tran > 2,]
head(Lcustomers); nrow(Lcustomers)
vector <- c(sum(Lcustomers$package), sum(Lcustomers$medium), sum(Lcustomers$small), sum(Lcustomers$large))
names <- c("package", "medium", "small", "large")
paste("Product type bought most is: ", names[which(vector==max(vector))])

Category: SCustomers (Sleeping Customers)¶

Scustomers <- users[users$time.last > 60,]
head(Scustomers); nrow(Scustomers)
vector <- c(sum(Scustomers$package), sum(Scustomers$medium), sum(Scustomers$small), sum(Scustomers$large))
names <- c("package", "medium", "small", "large")
paste("Product type bought most is: ", names[which(vector==max(vector))])

Category: ICustomers¶

AOVquantile <- quantile(users$Revenue/users$Tran)
AOVquantile

Icustomers <- users[users$Revenue/users$Tran > AOVquantile[4],]
head(Icustomers); nrow(Icustomers)
vector <- c(sum(Icustomers$package), sum(Icustomers$medium), sum(Icustomers$small), sum(Icustomers$large))
names <- c("package", "medium", "small", "large")
paste("Product type bought most is: ", names[which(vector==max(vector))])

These three categories of customers can be used to market better towards customers.

Clustering¶

An important point to note is that we shall not consider number of transactions as a feature for clustering because if a new customer comes, we shall not have number of transaction for him/her.

Considerations in clustering User IDs¶

This is a time series data
Every used ID can be considered to be following a time series with a certain length
The lengths of time series for every user is different

So, clustering here is equivalent to clustering several time series. A popular approach of clustering through K-means may not be efficient and may not converge.

Further consideration:¶

However, if we aggregate data by Cross Device user IDs and ignore the dates, K-means may provide efficient convergence. Further, since for a new customer we won't know the days anyway.

set.seed(123)
# Compute and plot wss for k = 2 to k = 15.
k.max <- 15
udata <- users[,-c(1,9,10)]
wss <- sapply(1:k.max, 
              function(k){kmeans(udata, k, nstart=50,iter.max = 15)$tot.withinss})
wss
plot(1:k.max, wss,
     type="b", pch = 19, frame = FALSE, 
     xlab="Number of clusters K",
     ylab="Total within-clusters sum of squares")

If we chose k=3, the total within-clusters explained variation won't be increased that much by increasing the size of clusters. Now, we will run k-means clustering using k=4

kmean.res <- kmeans(udata, 3)
kmean.res

K-means clustering with 3 clusters of sizes 63, 786, 534

Cluster means:
    package     medium     small      large     Tran   Revenue quantity
1 0.8412698 0.03174603 0.4444444 0.04761905 3.238095 141.04762 3.555556
2 0.3346056 0.23664122 0.4033079 0.03816794 1.262087  24.94911 1.298982
3 0.3389513 0.20411985 0.4531835 0.04119850 1.252809  24.02622 1.280899
  time.first time.last
1    39.0000  32.22222
2    48.1056  47.43384
3    16.6030  15.29963

Clustering vector:
   [1] 3 2 3 2 3 3 2 2 3 3 2 3 2 2 2 2 2 3 2 3 2 3 2 3 1 3 2 2 2 3 3 2 2 3 2 3 2
  [38] 3 2 1 2 3 2 3 1 3 1 2 2 2 3 2 2 1 3 2 2 2 2 2 2 2 2 2 3 2 2 2 3 2 3 2 2 1
  [75] 3 2 2 2 2 2 3 3 2 3 3 2 3 3 2 2 1 2 2 2 2 3 2 2 2 3 3 2 2 2 3 3 2 3 2 3 2
 [112] 3 3 2 2 2 3 2 2 1 3 3 3 3 2 2 2 3 2 3 2 2 3 2 1 2 2 3 2 3 3 3 2 2 3 2 2 3
 [149] 2 2 3 2 3 3 2 3 3 3 3 2 2 2 2 2 3 2 2 3 2 2 3 1 2 1 3 3 2 2 2 2 2 2 3 3 2
 [186] 2 3 2 2 3 3 3 2 2 3 2 2 3 3 2 2 2 1 3 2 1 3 3 2 1 3 2 2 3 3 2 2 2 2 2 2 2
 [223] 1 2 2 3 2 2 3 3 3 2 2 2 3 3 2 3 3 3 2 2 2 2 2 2 3 2 2 3 3 2 2 2 3 3 3 2 1
 [260] 2 3 3 2 1 3 2 3 2 3 2 3 3 3 2 2 3 3 3 1 2 3 2 3 2 2 2 2 2 1 2 3 2 2 3 2 2
 [297] 2 3 2 3 3 3 2 2 1 3 3 2 3 3 3 2 2 3 2 3 3 3 2 2 3 1 3 2 2 2 3 3 2 2 2 2 2
 [334] 3 2 3 3 2 2 3 2 2 2 2 2 2 3 2 1 3 2 2 2 2 3 2 2 2 2 2 2 1 3 2 2 2 3 3 2 2
 [371] 3 2 2 3 3 3 1 3 2 3 2 3 2 2 2 2 3 2 2 3 1 2 2 3 3 3 3 2 2 2 2 3 2 3 2 2 2
 [408] 2 2 3 2 2 3 2 3 3 2 3 2 2 2 3 3 2 3 2 2 3 3 2 2 2 3 3 3 2 2 2 2 2 2 3 3 2
 [445] 2 2 3 2 2 2 3 2 2 2 2 3 3 3 2 2 2 3 3 2 3 3 2 2 3 2 2 3 2 2 2 2 2 2 2 2 3
 [482] 3 2 2 3 2 3 2 1 3 2 2 2 2 3 3 3 2 3 3 3 2 2 3 3 3 2 2 2 2 3 2 2 3 2 3 2 2
 [519] 3 2 3 3 2 3 2 2 2 3 1 3 3 2 2 1 2 3 2 2 3 3 3 2 2 3 3 3 2 2 2 2 3 2 2 3 3
 [556] 2 2 3 2 3 3 3 3 3 3 3 2 2 1 3 2 2 2 2 2 2 2 2 2 2 2 2 3 2 2 2 3 3 2 3 2 2
 [593] 2 2 3 2 2 2 3 2 3 2 2 2 3 2 1 2 1 3 2 3 2 2 3 1 2 3 2 3 3 2 3 3 3 2 3 2 2
 [630] 3 3 2 2 2 2 2 2 3 2 2 2 3 1 2 2 3 2 2 3 3 3 2 3 2 3 3 2 3 3 3 3 2 2 2 2 3
 [667] 2 2 3 1 2 3 2 2 2 3 2 2 3 2 2 3 2 2 2 2 2 2 2 3 2 2 1 2 2 2 3 3 2 3 2 2 2
 [704] 2 3 2 3 2 2 2 2 3 2 2 2 2 2 3 2 2 3 2 2 3 2 2 2 2 3 2 2 3 2 2 2 2 3 3 3 2
 [741] 3 2 2 3 3 3 2 2 2 3 2 2 3 2 3 2 2 2 2 2 2 3 3 3 3 3 3 3 2 2 3 3 3 2 2 2 2
 [778] 3 3 2 3 3 2 3 2 3 3 3 2 3 3 2 2 2 3 2 3 2 2 3 3 2 3 1 2 3 3 2 3 2 2 2 3 3
 [815] 2 2 3 3 2 3 3 2 1 2 2 2 3 3 3 2 2 2 2 2 2 2 3 2 2 2 2 3 3 1 2 3 1 2 1 2 2
 [852] 2 3 3 2 2 3 2 3 2 2 2 2 2 2 2 3 2 3 3 2 3 3 2 2 3 3 3 2 2 2 3 2 2 3 2 3 3
 [889] 2 3 2 2 2 2 2 3 2 3 2 3 3 2 2 3 2 2 3 3 2 3 2 1 1 2 2 3 2 2 2 2 3 1 3 2 2
 [926] 3 3 3 3 3 2 2 3 3 2 3 3 3 3 2 2 2 3 2 2 2 2 3 2 2 2 3 2 2 2 3 2 2 2 3 2 2
 [963] 3 3 3 3 2 2 3 2 2 2 3 2 2 2 3 2 3 1 3 3 3 3 2 3 2 3 2 3 2 2 2 3 3 3 2 2 1
[1000] 2 2 2 3 2 3 2 3 2 2 2 2 3 2 3 3 2 2 2 1 2 2 2 2 2 3 2 3 3 3 2 1 3 3 2 2 2
[1037] 2 2 2 3 2 2 3 2 2 2 1 3 2 2 1 2 3 2 3 2 3 3 2 2 2 2 2 3 2 2 2 2 2 2 1 3 2
[1074] 2 2 2 2 2 3 2 1 2 2 3 2 3 2 2 2 3 3 3 3 2 2 3 3 2 2 3 2 2 3 2 3 2 3 3 3 2
[1111] 3 3 2 3 3 3 3 2 3 2 3 3 2 2 2 2 3 3 2 2 2 2 2 3 2 2 3 3 3 1 3 3 2 2 1 2 2
[1148] 2 3 2 3 2 2 2 2 2 3 2 2 2 3 2 2 3 2 3 3 2 3 3 3 2 2 2 2 2 3 2 2 2 3 2 2 3
[1185] 2 1 1 3 3 2 3 3 3 2 2 2 2 3 3 3 2 3 2 3 2 2 2 2 3 3 2 2 2 3 3 2 1 3 2 2 2
[1222] 2 2 3 2 3 3 3 3 2 3 2 3 2 2 2 3 2 3 2 3 2 2 2 2 1 2 3 2 3 3 2 2 2 2 3 2 3
[1259] 2 3 3 3 3 2 3 3 3 2 2 1 3 3 3 2 2 2 2 2 3 2 1 2 3 2 3 2 3 3 2 3 3 1 3 2 2
[1296] 2 3 2 2 3 2 2 2 3 1 3 3 2 2 2 2 3 3 2 2 3 2 3 2 3 3 2 3 2 3 3 2 3 3 2 2 2
[1333] 2 2 3 2 2 2 2 2 3 2 2 2 3 3 2 3 3 2 2 2 1 2 2 2 3 2 3 2 3 3 2 3 2 2 2 2 2
[1370] 2 3 2 2 3 2 2 3 3 3 2 1 2 3

Within cluster sum of squares by cluster:
[1] 356193.5 346187.0 232988.1
 (between_SS / total_SS =  61.0 %)

Available components:

[1] "cluster"      "centers"      "totss"        "withinss"     "tot.withinss"
[6] "betweenss"    "size"         "iter"         "ifault"

km.res$cluster <- as.factor(kmean.res$cluster)
clusteredusers <- users
clusteredusers$cluster <- kmean.res$cluster
head(clusteredusers)
paste("Size of cluster", c(1,2,3), ":", summary(as.factor(clusteredusers$cluster)))

summary(clusteredusers[clusteredusers$cluster==1,])
summary(clusteredusers[clusteredusers$cluster==2,])
summary(clusteredusers[clusteredusers$cluster==3,])

           Cross.device.user.ID    package           medium       
 039df8d15005b9008876: 1        Min.   :0.0000   Min.   :0.00000  
 0541abca3ea8a78ce2d8: 1        1st Qu.:0.0000   1st Qu.:0.00000  
 072a0c5c702131eccff9: 1        Median :0.0000   Median :0.00000  
 076e8674be4d2ad37a8d: 1        Mean   :0.8413   Mean   :0.03175  
 0914ddf25bea3630e32b: 1        3rd Qu.:2.0000   3rd Qu.:0.00000  
 0d2369ec05e610947ed1: 1        Max.   :5.0000   Max.   :2.00000  
 (Other)             :57                                          
     small            large              Tran          Revenue   
 Min.   :0.0000   Min.   :0.00000   Min.   :1.000   Min.   : 83  
 1st Qu.:0.0000   1st Qu.:0.00000   1st Qu.:2.000   1st Qu.:101  
 Median :0.0000   Median :0.00000   Median :3.000   Median :115  
 Mean   :0.4444   Mean   :0.04762   Mean   :3.238   Mean   :141  
 3rd Qu.:0.0000   3rd Qu.:0.00000   3rd Qu.:4.000   3rd Qu.:161  
 Max.   :7.0000   Max.   :3.00000   Max.   :8.000   Max.   :559  
                                                                 
    quantity       firstdate             lastdate            time.first  
 Min.   :1.000   Min.   :2017-02-13   Min.   :2017-02-14   Min.   : 3.0  
 1st Qu.:2.000   1st Qu.:2017-02-25   1st Qu.:2017-02-27   1st Qu.:26.5  
 Median :3.000   Median :2017-02-28   Median :2017-03-16   Median :47.0  
 Mean   :3.556   Mean   :2017-03-08   Mean   :2017-03-14   Mean   :39.0  
 3rd Qu.:4.000   3rd Qu.:2017-03-20   3rd Qu.:2017-03-30   3rd Qu.:50.0  
 Max.   :8.000   Max.   :2017-04-13   Max.   :2017-04-13   Max.   :62.0  
                                                                         
   time.last        cluster 
 Min.   : 3.00   Min.   :1  
 1st Qu.:16.50   1st Qu.:1  
 Median :31.00   Median :1  
 Mean   :32.22   Mean   :1  
 3rd Qu.:48.00   3rd Qu.:1  
 Max.   :61.00   Max.   :1

           Cross.device.user.ID    package           medium      
 003d60169becee4e390d:  1       Min.   :0.0000   Min.   :0.0000  
 00b1503d497a961b919e:  1       1st Qu.:0.0000   1st Qu.:0.0000  
 011ba287dd801f3f98c0:  1       Median :0.0000   Median :0.0000  
 013c56f4157b7cd6d644:  1       Mean   :0.3346   Mean   :0.2366  
 017cbe8dcd707a1fe1dc:  1       3rd Qu.:1.0000   3rd Qu.:0.0000  
 01c32cefc97342474024:  1       Max.   :4.0000   Max.   :4.0000  
 (Other)             :780                                        
     small            large              Tran          Revenue     
 Min.   :0.0000   Min.   :0.00000   Min.   :1.000   Min.   : 0.00  
 1st Qu.:0.0000   1st Qu.:0.00000   1st Qu.:1.000   1st Qu.:13.00  
 Median :0.0000   Median :0.00000   Median :1.000   Median :19.00  
 Mean   :0.4033   Mean   :0.03817   Mean   :1.262   Mean   :24.95  
 3rd Qu.:1.0000   3rd Qu.:0.00000   3rd Qu.:1.000   3rd Qu.:33.00  
 Max.   :4.0000   Max.   :2.00000   Max.   :4.000   Max.   :84.00  
                                                                   
    quantity       firstdate             lastdate            time.first   
 Min.   :1.000   Min.   :2017-02-13   Min.   :2017-02-13   Min.   :32.00  
 1st Qu.:1.000   1st Qu.:2017-02-20   1st Qu.:2017-02-20   1st Qu.:42.00  
 Median :1.000   Median :2017-02-27   Median :2017-02-27   Median :48.00  
 Mean   :1.299   Mean   :2017-02-26   Mean   :2017-02-27   Mean   :48.11  
 3rd Qu.:1.000   3rd Qu.:2017-03-05   3rd Qu.:2017-03-06   3rd Qu.:55.00  
 Max.   :7.000   Max.   :2017-03-15   Max.   :2017-04-04   Max.   :62.00  
                                                                          
   time.last        cluster 
 Min.   :12.00   Min.   :2  
 1st Qu.:41.00   1st Qu.:2  
 Median :48.00   Median :2  
 Mean   :47.43   Mean   :2  
 3rd Qu.:55.00   3rd Qu.:2  
 Max.   :62.00   Max.   :2

           Cross.device.user.ID    package          medium      
 00146fb6a0c24b74232d:  1       Min.   :0.000   Min.   :0.0000  
 0086c468fd6bc0d2ff61:  1       1st Qu.:0.000   1st Qu.:0.0000  
 00de87f3528f3597af00:  1       Median :0.000   Median :0.0000  
 00eeb91a2c9ff36766a4:  1       Mean   :0.339   Mean   :0.2041  
 014a04bf9674f4ee0032:  1       3rd Qu.:1.000   3rd Qu.:0.0000  
 0171a55cdca297aca753:  1       Max.   :2.000   Max.   :3.0000  
 (Other)             :528                                       
     small            large             Tran          Revenue     
 Min.   :0.0000   Min.   :0.0000   Min.   :1.000   Min.   : 0.00  
 1st Qu.:0.0000   1st Qu.:0.0000   1st Qu.:1.000   1st Qu.:13.00  
 Median :0.0000   Median :0.0000   Median :1.000   Median :19.00  
 Mean   :0.4532   Mean   :0.0412   Mean   :1.253   Mean   :24.03  
 3rd Qu.:1.0000   3rd Qu.:0.0000   3rd Qu.:1.000   3rd Qu.:30.00  
 Max.   :3.0000   Max.   :2.0000   Max.   :6.000   Max.   :88.00  
                                                                  
    quantity       firstdate             lastdate            time.first   
 Min.   :1.000   Min.   :2017-02-15   Min.   :2017-03-15   Min.   : 0.00  
 1st Qu.:1.000   1st Qu.:2017-03-22   1st Qu.:2017-03-23   1st Qu.: 8.00  
 Median :1.000   Median :2017-03-31   Median :2017-04-02   Median :16.00  
 Mean   :1.281   Mean   :2017-03-30   Mean   :2017-03-31   Mean   :16.60  
 3rd Qu.:1.000   3rd Qu.:2017-04-08   3rd Qu.:2017-04-09   3rd Qu.:24.75  
 Max.   :6.000   Max.   :2017-04-16   Max.   :2017-04-16   Max.   :60.00  
                                                                          
   time.last       cluster 
 Min.   : 0.0   Min.   :3  
 1st Qu.: 7.0   1st Qu.:3  
 Median :14.0   Median :3  
 Mean   :15.3   Mean   :3  
 3rd Qu.:24.0   3rd Qu.:3  
 Max.   :32.0   Max.   :3

We checked if the clusters are dissimilar and if customers in every cluster are similar.

Cluster 1 consists of customers who buy on average 3 products (median) and their median expenditure is 43 Euros.
Cluster 2 consists of customers who buy on average 1 product (median) and their median expenditure is 19 Euros.
Cluster 2 consiste of customers who buy on average 1 product (median) and their median expenditure is around 19 Euros.

The difference between customers in cluster 2 and cluster 3 lies in the types of product they bought. In cluster 2, more people buy small and large items while in cluster 1, more people buy packages and medium items.

New Customers¶

Now we'll identify frequency of new clients by month

# Get data for sales and visits
newusersdata <- data[!data$Existing.client %in% 0,]
newusers <- newusersdata %>%
    group_by(YearMonth) %>%
    summarise(Existing.client = sum(Existing.client), Sales = sum(Dummy, na.rm=TRUE)) %>%
    mutate(Salesratio = round(Sales/Existing.client, 2)) %>%
    arrange(desc(Existing.client))
newusers

It is clear that during the months when revenue was the highest, more new users joined in. One thing to note is the sales ratio that is the ratio of Number of sales to newusers and their number of visits is less. One idea will be not only gain new clients but also increase this sales ratio since this sales ratio is fairly average over the year with a standard deviation of less than 0.03

Preliminary Recommendations: My recommendation is to focus on getting more customers who did not buy but just visisted to the website of this retailer. My other recommendation would be to focus on Applications as they bring more transactions. Lastly, it is important to conduct customer centric marketing as there are many customers who purchased only once.¶

Date	Rev.
2017-04-14 14:43:44	427
2017-03-31 16:31:55	425
2017-01-21 04:20:05	320
2016-12-30 14:24:41	317
2017-02-18 12:45:25	293

Date	Tra
2016-12-10 21:19:01	7
2016-06-27 01:38:55	6
2016-07-04 00:26:55	6
2016-11-02 12:07:09	6
2017-01-19 22:15:32	6
2017-01-21 04:20:05	6
2017-01-30 17:36:42	6

YearMonth	Rev.	Percentage
2016 Dec	88297	19.61
2016 Nov	60391	13.41
2017 Feb	47338	10.51
2017 Jan	38067	8.45
2016 Sep	37016	8.22
2016 Oct	35831	7.96
2016 Apr	31927	7.09
2017 Mar	29050	6.45
2016 May	24030	5.34
2016 Aug	21807	4.84
2016 Jun	19323	4.29
2016 Jul	17169	3.81

YearMonth	Rev.	Tran	AOV	Percentage
2016 Dec	88297	4065	21.72	18.68
2016 Nov	60391	2563	23.56	11.78
2017 Feb	47338	2201	21.51	10.12
2017 Jan	38067	1827	20.84	8.40
2016 Sep	37016	1984	18.66	9.12
2016 Oct	35831	1668	21.48	7.67
2016 Apr	31927	1570	20.34	7.22
2017 Mar	29050	1347	21.57	6.19
2016 May	24030	1306	18.40	6.00
2016 Aug	21807	1101	19.81	5.06
2016 Jun	19323	1025	18.85	4.71
2016 Jul	17169	1099	15.62	5.05

YearMonth	Visits	sales	svisits
2016 Dec	41799	4065	0.10
2016 Apr	16747	1570	0.09
2016 Nov	27295	2563	0.09
2017 Feb	24607	2201	0.09
2017 Jan	23306	1827	0.08
2017 Mar	16861	1347	0.08
2016 Jul	15979	1099	0.07
2016 Oct	23700	1668	0.07
2016 Sep	27979	1984	0.07
2016 May	21312	1306	0.06
2016 Jun	18673	1025	0.05
2016 Aug	25469	1101	0.04

Device	Revenue	Tran	AOV	Revenue.pct	svisits	ABV	Product.visits
Desktop	359827	16314	22.06	76.92	0.08	22.72	209837
Mobile	102373	5900	17.35	21.88	0.09	17.93	67624
Tablet	5602	324	17.29	1.20	0.02	19.46	15774

Cross.device.user.ID	package	medium	small	Tran	Revenue	quantity	firstdate	lastdate	time.first	time.last
00146fb6a0c24b74232d	1	0	0	1	15	1	2017-03-25	2017-03-25	22	22
003d60169becee4e390d	0	1	0	1	17	1	2017-03-03	2017-03-03	44	44
0086c468fd6bc0d2ff61	0	0	1	1	10	1	2017-03-15	2017-03-15	32	32
00b1503d497a961b919e	0	0	1	1	28	1	2017-03-08	2017-03-08	39	39
00de87f3528f3597af00	0	0	1	1	23	1	2017-03-22	2017-03-22	25	25
00eeb91a2c9ff36766a4	1	0	0	1	57	1	2017-04-09	2017-04-09	7	7

Cross.device.user.ID	small	Tran	Revenue	quantity	firstdate	lastdate	time.first	time.last
02a7b34386ba592c1ef6	3	3	21	3	2017-04-01	2017-04-05	15	11
0373ee30ba2f1f1ca8e4	0	3	29	3	2017-03-14	2017-03-16	33	31
039df8d15005b9008876	0	4	93	4	2017-02-25	2017-02-25	50	50
03a6bc25103535721051	0	4	67	5	2017-02-22	2017-04-09	53	7
0541abca3ea8a78ce2d8	0	3	144	3	2017-04-04	2017-04-09	12	7
057bb52d2a6c777bfad3	0	3	50	3	2017-02-14	2017-02-28	61	47

Cross.device.user.ID	package	medium	small	large	Tran	Revenue	quantity	firstdate	lastdate	time.first	time.last
03192220e87f8e23b5e5	0	0	1	0	1	9	1	2017-02-14	2017-02-14	61	61
043aff833259514934ab	0	0	0	1	1	18	1	2017-02-13	2017-02-13	62	62
131350b39be3338fea43	0	1	0	0	1	9	1	2017-02-13	2017-02-13	62	62
13d56f9d9e836a10ae7f	0	0	1	0	1	7	1	2017-02-14	2017-02-14	61	61
14397d06bcab24fd7d8a	0	0	1	0	1	6	1	2017-02-13	2017-02-13	62	62
1ec74f90e0bd3b97414d	1	0	0	0	1	21	1	2017-02-13	2017-02-13	62	62

YearMonth	Existing.client	Sales	Salesratio
2016 Dec	6810	830	0.12
2016 Nov	5606	751	0.13
2017 Feb	4444	554	0.12
2017 Jan	3980	455	0.11
2016 Sep	3947	391	0.10
2016 Apr	3629	574	0.16
2016 Oct	3506	408	0.12
2017 Mar	3147	328	0.10
2016 Aug	2592	199	0.08
2016 May	2508	253	0.10
2016 Jun	2269	193	0.09
2016 Jul	2193	205	0.09
2017 Apr	1913	172	0.09
2016 Mar	280	35	0.12